Authors :
First Author: Sungeun Hwang, MD – Ewha Womans University Mokdong Hospital
Presenting Author: Kyung-Il Park, MD, PhD – Seoul National University Hospital Healthcare system Gangnam center
Youmin Shin, MS – Seoul National University Hospital; Jun-Sang Sunwoo, MD/PhD – Kangbuk Samsung Hospital; Hyoshin Son, MD/PhD – Catholic University of Korea; Seung-Bo Lee, PhD – Keimyung University School of Medicine; Kon Chu, MD/PhD – Seoul National University Hospital; Ki-Young Jung, MD/PhD – Seoul National University Hospital; Sang Kun Lee, MD/PhD – Seoul National University Hospital; Young-Gon Kim, PhD – Seoul National University Hospital; Kyung-Il Park, MD/PhD – Seoul National University Hospital Healthcare System Gangnam Center
Rationale:
While 60-70% of patients with newly diagnosed epilepsy become seizure-free with proper antiseizure medication (ASM) treatment, the remaining 30-40% of patients are drug-resistant. Temporal lobe epilepsy (TLE) is the most common type of focal epilepsy and is frequently drug-resistant. Surgical treatment such as anterior temporal lobectomy is an effective treatment option for drug-resistant TLE with success rates of 60-70%. However, surgical treatment tends to be reserved as a last resort and drug-resistant epilepsy and this tendency may delay timely surgical treatment. Therefore, early identification of drug-resistant epilepsy can minimize unnecessary seizure relapses. Although several studies predicted surgical outcome based on electroencephalography (EEG), there is no study predicting drug response in the early stages of monotherapy. In this study, we analyzed resting-state EEG using feature-based machine-learning algorithms to predict drug responsiveness in TLE.
Methods:
Fifty-two patients with TLE whose EEGs were acquired during monotherapy at Seoul National University Hospital and Kangbuk Samsung Hospital between 2014 and 2021 were included in this study. All EEG data were acquired using the modified international 10-20 electrode placement system at a sampling rate of 250 Hz. EEG were epoched to a time window of 120 seconds with 50% overlap. Various features (Hjorth, statistical, energy, zero-crossing, correlation, coherence) were extracted. Optimal frequency for each feature was evaluated using random forest (RF) model. For the optimal feature and frequency, the classification was performed using RF, extreme gradient boosting (XGB), and light gradient boosting (LGB) models. Features were selected with filter methods (chi-square, ANOVA, mutual information) and wrapper method (recursive feature elimination).
Results:
Thirty-eight patients (73%) were drug-responsive and 14 patients (27%) were drug-resistant for follow-up periods of 37.9 ± 21.9 months. Interictal epileptic discharges were present in EEG data of 27 patients. Among 6 types of features, correlation showed the best performance in gamma frequency band and was used for further experiment. RF model with ANOVA-based filter method feature selection showed the highest performance (AUC = 0.752, accuracy = 0.713, positive predictive value = 0.743, negative predictive value = 0.730, Figure 1). The selected features, which include connectivity channels involving the frontal and central areas, are important for prediction.
Conclusions:
This study shows that acceptable performance can be achieved with optimal feature and frequency band selection. This can be applied for early identification of drug-resistant TLE, which is a potential candidate for surgical treatment.
Funding:
This study was funded by the National Center for Mental Health (grant number: NHER22A01), Republic of Korea and Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HI22C0776).